Knowledge Distillation for Efficient Sequences of Training Runs